An adaptive deep-learning load forecasting framework by integrating transformer and domain knowledge

• DOI: 10.1016/j.adapen.2023.100142
• Publication Date: 2023-05-08T05:55:49Z
• Authors: Jiaxin Gao; Yuntian Chen; Wenbo Hu; Dongxiao Zhang
• Keywords: Interpretability; Adaptive Learning; Robustness

Electrical energy is essential in today's society. Accurate electrical load forecasting beneficial for better scheduling of electricity generation and saving energy. In this paper, we propose an adaptive deep-learning framework by integrating Transformer domain knowledge (Adaptive-TgDLF). Adaptive-TgDLF introduces the model learning methods (including transfer different locations online time periods), which captures long-term dependency series, more appropriate realistic scenarios with scarce samples variable data distributions. Under theory-guided framework, divided into dimensionless trends local fluctuations. The are considered as inherent pattern load, fluctuations to be determined external driving forces. Adaptive can cope change location time, make full use at times train a efficient model. Cross-validation experiments on districts show that approximately 16% accurate than previous TgDLF saves half training time. 50% weather noise has same accuracy without noise, proves its robustness. We also preliminarily mine interpretability Adaptive-TgDLF, may provide future potential theory guidance. Furthermore, demonstrate accelerate convergence number epochs achieve performance, enables results changing load.